The Fascination With Robots Folding Clothes Explained

It seems like every week there's a new video of a robot folding clothes. We've had some fantastic demos such as this is a semi-autonomous video from Weave Robotics on X.

It's amazing stuff, but Weave isn't the only company making videos like this. Figure 02 – folding clothes.. Figure 03 – folding clothes.. Physical intelligence launched its flagship vision-language-action model pi0 with amazing video of a robot folding clothes after unloading the washing machine. You can see robots folding clothes. live at robotics exhibitions. Even before all this, Google showed clothes folding at work, ALOHA released. 7X Tech even plans to sell robots for folding clothes!

Besides folding real clothes, there are other tasks that are similar to folding clothes, such as Dayna's task. folding napkin– which leads to what is probably my best Robot Video of the Yeardemonstrating 18 hours of continuous napkin folding. So why all this robotic manipulation Do companies suddenly fold?

Reason 1. Previously, we practically could not do this.

Got a job going back more than a decade this shows how some robotic clothing is folded. But these demos were extremely unstable, extremely slow, and not even remotely ready for production. Previous solutions existed (even learning-based solutions!), but they relied on precise camera calibration or carefully hand-crafted features, which meant that these clothing folding demonstrations typically only worked on one robot, in one environment, and perhaps only worked once—just long enough to record a demo video or present a paper.

With a little help from a creative patterned shirt, PR2 I was putting things away back in 2014.Bosch/IEEE

Take a look at this example PR2 UC Berkeley folding underwear since 2014. This robot actually uses a neural network policy. But this policy is very small and fragile; it selects and places objects on the same green background, moves very slowly, and can't handle a wide variety of shirts. Putting this work into practice will require larger models pre-trained on web-scale data, as well as better, more general imitation learning methods.

And now, 10 years later, with the availability of appropriate demonstration data, many different startups and research laboratories were able to implement clothing folding demonstration models; this is what we have seen in many amateurs and startups using largely similar tools (e.g. LeRobot from HuggingFace), without narrow specialization.

Reason 2: It looks great and people want it!

Many of us who work in robotics having that “north star” robot butler that can do all the housework we don't want to do. Mention folding clothes and many, many people will sign that they will never want to fold clothes again and will be willing to part with almost any amount of money to make it happen.

This is also important for the participating companies. Companies like Fig and 1x are raising large sums of money based on the idea that they will be able to automate many different jobs, but these companies increasingly seem to want to start from home.

Dyna Robotics can fold an indefinite number of napkins indefinitely.Daina Robotics

And that's part of the magic of these demos. Although they are slow and imperfect, everyone can begin to imagine this technology becoming what we all want: a robot that can exist in our home and alleviate all those everyday hassles that take up our time.

Reason 3: This avoids what robots are still bad at.

These robotic behaviors are generated by models trained through imitation learning. Modern methods of teaching imitation such as Distribution Policy use methods inspired generative AI produce complex deft robot trajectories based on examples of expert human behavior provided to them, and they often require many, many trajectories. Job ALOHA on the loose from Google is a great example: it takes about 6,000 demonstrations to learn, for example, how to tie your shoes. For each of these demonstrations, a person piloted a pair robotic arms during the execution of a task; all this data was then used to train the policy.

We need to remember what is difficult about these demonstrations. Human demonstrations never idealnor are they entirely consistent; for example, two human demonstrators will never grab accurate the same part of an object with millimeter precision. This is potentially an issue if you want to screw the lid onto the top of a machine you're building, but it's not a problem at all for folding clothes, which is quite forgiving. This has two side effects:

• This makes it easier to assemble the clothes folding demos you need, since you don't have to throw out every practice demo that's a millimeter out of specification.
• You can use cheaper, less repeatable hardware to perform the same task, which is useful if you suddenly need a fleet of robots collecting thousands of demos, or if you have a small team with limited funding!

For the same reasons, it's great that by folding the fabric you can repair cameras in the correct position. When learning a new skill, you need training examples that “cover” the environment space you expect to see during deployment. So, the more control you have, the more effective your learning process will be—the less data you'll need, and the easier it will be to deliver a compelling demo. Keep this in mind when you see a robot stacking things on a plain tabletop or against a very clean background; This is not just a beautiful frame, it really helps the robot!

And because we've committed to collecting a ton of data—tens of hours—to make this task work well, mistakes will be made. It's very useful if it's easy reboot task, i.e. restore it to a state from which you can try to perform the task again. If something goes wrong with folding clothes, it's okay. Just take the fabric, throw it, and you're ready to start over. This won't work if, say, you're stacking glasses in a cupboard, because if you knock over the stack or drop one on the floor, you'll be in trouble.

Folding your clothes also allows you to avoid too much contact with the environment. Once you apply a lot of pressure, things can break down, the task can become irreversible, and demonstrations are often much more difficult to assemble because the forces are not as easy to observe politics. And every change (such as the amount of force you apply) will eventually require more data in order for the model to “cover” the space in which it needs to operate.

What to expect

Although we see a lot of demonstrations of folding clothes these days, many of them still impress me. As mentioned above, Dyna was one of my favorite demos this year, mainly because long-term robot policies have been rare so far. But they were able to demonstrate folding without additional training data (that is, folding without additional training data) at several different conferences, including Take action in San Francisco And Conference on Robot Learning (CoRL) in Seoul. This is impressive and actually very rare in robotics even now.

In the future, we should hope to see robots that can handle more complex and dynamic interactions with the environment: moving faster, moving heavier objects, climbing, or otherwise handling difficult terrain while performing manipulative tasks.

But for now, remember that modern teaching methods have their strengths and weaknesses. It seems that folding clothes, while not easy, is a task that is very well suited to what our models can do right now. So expect to see a lot more of this.